The goal of the present contribution is to shed light on the textual properties of written outputs generated by large language models by focusing on the referential dimension of textual organization. To gain insights on this aspect, we analyze the properties of the referring expressions forming the main referential chain running through biographies, which typically correspond to the personality on which the text is centered. Based on an empirical qualitative corpus-driven analysis of 30 biographies generated in Italian by GPT-4, we describe (i) the forms of the linguistic expressions codifying the discourse referents building the main referential chain of the biographies; (ii) the degree of complexity of these linguistic expressions, verifying if they match the degree of cognitive accessibility of the discourse referents; finally, on a textual level, (iii) the architectures that these chains form and the distribution of the chain’s rings in the textual units composing the biography. Our analysis reveals that the main referential chain building the analyzed biographies is generally well-formed but very simple and relies on repetitive textual patterns. At a micro-textual level, we find marked textual patterns such as the over-codification of a discourse referent as well as cases of over-segmentation of semantically and pragmatically compact textual units.